Estimating Bathymetric Depth with Satellite Images: The Case of California

ABSTRACT

Bathymetry is the general name given to the studies of topographic measurement of the water body base. One of the most important data obtained as a result of these measurements is the water depth data. Accurate water depth data is required for studies of the sea, coastal ecosystems, construction of coastal areas, ports and infrastructures, biological oceanography, sea level rise etc. The classical sounding method, one of the bathymetric methods used since ancient times, has left its place to a number of methods such as acoustic sonar systems, Laser (Light Amplification by Stimulated Emission of Radiation) and LIDAR (Laser Imaging Detection and Ranging) systems. In recent years, many studies have been carried out to obtain bathymetric data in shallow waters using satellite remote sensing data. In this study, the California coastline, for which bathymetric data was obtained before, was preferred as the test area. Landsat-8 satellite images of the California coastal region were processed using QGIS 3. 16 and ArcMap 10.5 geographic information system software using various band combinations, and depth data were obtained with an open source GUI (Graphical User Interface) written in Python programming language. K Nearest Neighborhood method, which is a machine learning algorithm, was used in the depth estimation process. The obtained data were compared with the reference depth data derived from the current digital elevation model of the California coastline, and RMSE (Rood Mean Square Error) error rates were determined for each band combination. The results show that the bathymetric accuracy is good and the RMSE is less than 20% of the maximum depth in both study areas.

1. MATERIALS

Reference Digital Elevation Model (DEM)

National Geophysical Data Center (NGDC) creates high-resolution digital elevation models (DEMs) for specific U.S. coastal areas. DEM data (Figure 1.1) used as a reference in this study was provided by NGDC in 2010, in WGS84 horizontal datum and 30 meters spatial resolution. [1]

Figure 1.1 California Coastline DEM Data

Satellite Images

In the selection of the satellite image to be used in the study, four spectral bands of the Landsat 8 satellite OLI sensor with a spatial resolution of 30 m, a radiometric resolution of 16, and a spectral range of 0.43 μm to 0.67 μm were used, since it is open source and free of charge. The download of spectral bands was performed according to the Landsat Collection 2 Level-2 option. Thanks to this feature, which provides the user with a minimum pre-processing requirement, there is no need to perform Atmospheric Correction manually. In the literature review, it has been determined that the bands that are frequently used for bathymetry studies are coastal aerosol and RGB bands. Therefore, in this study, combinations with the band types listed are included.

Spectral bands were obtained free of charge from the United States Geological Survey (USGS) site.

Table 1.1 Landsat 8 OLI Band Combinations [2]

Figure 1.2 Workflow [3]

2. DATA PREPROCESSING

Combining Spectral Bands

In the study, the most suitable 1, 2, 3 and 4 bands were selected from the 11-band multispectral data set in the Landsat 8 satellite image obtained from the USGS website. Two different combinations were used to see the effects of band combinations in bathymetric studies. These; 4-3-2 RGB and 4-3-1 RGC are color images.

Merging operations can be done by various resampling methods. These methods are nearest neighbor method, bilinear interpolation, cubic interpolation, averaging and mode methods. In the study, the nearest neighbor method was preferred as it is the most widely used method.

Creating a Mosaic Image

As a result of combining the spectral bands, 3 full-frame Landsat 8 satellite images covering the entire study area were obtained. In order to continue the work on an uninterrupted image, it is necessary to combine 3 full-frame images. As a result of the mosaic process, a total of 2 mosaic images, RGB and RGC images, were produced. (Figure 2.1)

Determining the Study Area

The maximum depth that sunlight can reach in water is known as about 30 meters. Since the depth data will be produced from satellite images in this study area, the area where the depth varies between 0 meters and -30 meters has been determined as the study area.

3. DATA PROCESSİNG

The data processing step is the stage in which the estimated depth values are obtained from the mosaic image by using the values in the depth samples.

Before making an estimation, the data set consisting of depth samples should be divided into two as training and test data sets. When the training dataset is processed, the SDB GUI is automatically divided into two as the training dataset and the validation dataset.

Training Model

Within the SDB GUI, there are four machine learning methods for depth estimation. The methods are K-Nearest Neighbors, Multiple Linear Regression, Random Forest, and Support Vector Machines. They are all methods using the Scikit-Learn module, which is a machine learning library.

In this study, the K-Nearest Neighbors method was preferred. This method implements learning based on k nearest neighbors of each query point. Within the scope of the study, a total of 8 trainings were carried out by applying various combinations of the number of neighbors and training and validation dataset size parameters, which are the parameters that can affect the depth estimation process.

Table 3.1 Comparison of Training Results

When the RMSE rates of the 8 trainings in total are compared, the training with the highest accuracy; It is seen that the training is carried out with the KNN algorithm with the coefficient K = 5 on the RGB image and the ratio of the training data set to the entire data set is determined as 90%. In Figure 3.1, the test data set containing the reference depth value and the estimation depth values were compared. Below is the visualization of the RGB image with the output raster DEM data as a result of the training in Figure 3.2. A hillshade pad was used to make the depth values meaningful.

Figure 3.1 Depth Estimation Errors Distribution Chart

4. RESULTS

As a result of the analyzes carried out, it was observed that the algorithm did not work well in some of the stream beds formed from the California coastline into the city. It has been noticed that in some parts of the coastline, depth values are also created on the land. This error is thought to be eliminated if satellite images with better spatial and radiometric resolution are studied. At the same time, the effect of fluctuations that may occur on the water surface is another source of error.

It is envisaged that this study can be improved by increasing the diversity of data sets and incorporating different algorithms. As a result, the applicability of the method, especially in regions with shallow depths, has been seen, and a basis for new studies on this subject has been established..

LITERATURE

[1] NOAA, NCEI, Coastal Elevation Models. https://www.ngdc.noaa.gov/mgg/coastal/coastal.html

[2] ArcGIS Blog, Band Combinations for Landsat 8 (2013). https://www.esri.com/arcgis-blog/products/product/imagery/band-combinations-for-landsat-8/

[3] Rifqi Muhammad Harrys (2021), Satellite Derived Bathymetry (SDB) GUI. GitHub.

[4] Earth Explorer, USGS, https://earthexplorer.usgs.gov

Geomatics Insight

by Bâlâhun Kalkan